askrene: don't crash if a layer is removed while getroutes waits

[asoltys]

getroutes checks its layer names when the command arrives, but apply_layers() only looks them up again once a child is free. If a layer is removed in between, find_layer() returns NULL and we fall through to:

cln-askrene: plugins/askrene/askrene.c:406: apply_layers: Assertion `streq(layernames[i], "auto.sourcefree")' failed.

The plugin is not restartable mid-call, so lightningd exits with it.

We hit this in production about 16 times in 3 hours during heavy xpay load. A fee-quoting path creates a temporary layer, calls getroutes, and removes the layer in a finally when the call returns or the client-side RPC timeout fires. With all askrene-max-threads children busy, the request was still in the waiters queue when the cleanup removed its layer, and the next dequeue took the node down.

This makes apply_layers() report the missing layer, so do_getroutes() fails just that request with JSONRPC2_INVALID_PARAMS ("Layer X was removed before routing began"). The early return happens before gossmap_apply_localmods(), and the partially built localmods and layers are allocated off cmd, so there's nothing to unwind.

The new test test_layer_removed_while_waiting sets askrene-max-threads to 0 so the request is guaranteed to queue, removes the layer, then sets it back to 1 and runs another request to wake the queue. Without the fix it reproduces the assert above; with it, it passes. The full tests/test_askrene.py passes locally (38 tests).

Checklist

• The changelog has been updated in the relevant commit(s) according to the guidelines.
• Tests have been added or modified to reflect the changes.
• Documentation has been reviewed and updated as needed. (no doc change: same error path as an unknown layer)
• Related issues have been listed and linked, including any that this PR closes. (none found)
• Important All PRs must consider how to reverse any persistent changes for tools/lightning-downgrade (no persistent changes)

🤖 Generated with Claude Code